后台按「运营 / 知识 / 系统」三组组织全部 18 个频道。从单人长期运营，到知识沉淀、AI 总控、请示裁决、自我演化与成效复盘——每个频道都对应真实可用的后端能力与可审计的数据。
The console organizes all 18 channels into three groups — Operations, Knowledge, System. From per-person long-term operations to knowledge curation, AI command, escalation, self-evolution and outcome review — every channel maps to real backend capability and auditable data.
围绕单个好友长期运营：好友池、单人画像、长期运营记忆、运营大脑、方法论、提示词与执行边界。会话可回放，可对单人发起模拟验证后再上线。
Long-term operations around a single contact: contact pool, per-person profile, long-term memory, operating brain, playbooks, prompts and execution bounds. Replay conversations; simulate before going live.
用一个后台管理 Agent 以自然语言统筹整个系统：查询、配置、发消息、建任务。展示执行计划、工具调用状态、Dry-run 演练，高风险动作需二次确认。
A management agent runs the whole system in natural language: query, configure, send, schedule. It shows execution plans, tool-call states, dry-run rehearsals; high-risk actions need confirmation.
运行态势总览：微信账号、运营对象、跟进任务与最近事件的整体状态，一屏掌握系统在做什么。
Operational overview: accounts, operated contacts, follow-up tasks and recent events at a glance — what the system is doing right now.
所有需要决策 / 审核的事项收口于此：请示裁决、知识核验、画像发布、经验晋升等 8 类待办经同一数据源汇聚，逐条处置、可深链直达源对象。
Everything needing a decision or review converges here: escalations, knowledge verification, profile publishing, lesson promotion — 8 streams from one data source, handled item by item with deep links.
配置幕后决策人链、触发请示的情形、超时转备选与推送频控。保存后即时生效于私聊运营域——这是「请示而非接管」的策略中枢。
Configure the decider chain, the situations that trigger escalation, timeout fallback and push throttling. Effective immediately in private-chat ops — the control hub for "consult, never hand off."
维护产品目录与价格，登记核实成交，查看客户当前持有与售后状态。产品声明必须有此处与知识库的可核验依据，否则发送被拦截。
Maintain product catalog and pricing, register verified deals, track what each customer holds and after-sales state. Product claims must be backed here and in the knowledge base, or the send is blocked.
群画像、群节奏与群工具工作流，作为独立运营域在下一阶段建设。当前为占位频道，不与私聊用户运营混用同一代码路径。
Group profiles, cadence and group-tool workflows are a separate operation domain planned next. Currently a placeholder — it shares no code path with private-chat user ops.
朋友圈内容计划、发布队列与互动复盘，同样作为独立运营域在下一阶段建设。当前为占位频道。
Moments content planning, publish queue and engagement review — also a separate domain planned next. Currently a placeholder.
聚焦 AI 在为这个人做什么决策、依据是什么、下一步最优动作如何排序，把长期关系经营讲清楚。
Focuses on what the AI is deciding for this person, on what grounds, and how next-best actions rank — making long-term relationship work legible.
直接暴露画像字段、记忆条目、提示词与边界配置，让熟悉系统的运营者快速核对与微调。
Exposes profile fields, memory entries, prompts and bounds directly, so seasoned operators can verify and fine-tune fast.
管理知识库领域 schema、缺口信号与切片修订历史。9 类知识切片互相编织成知识仓；每次修订双写——不可变修订流 + 可变当前切片，带 sha256 校验。AI 永不自动核验，新知识一律落草稿待审。
Manage domain schema, gap signals and revision history. 9 wiki types weave into a linked store; each revision double-writes — an immutable revision stream plus the mutable current slice, sha256-checked. The AI never self-verifies; new knowledge lands as draft for review.
维护产品资料、FAQ、话术、禁用表达、品牌语气与朋友圈素材，供 Agent 自主运营时调用。素材与知识库口径需保持一致，避免价格 / 政策冲突。
Maintain product docs, FAQs, scripts, banned phrasings, brand tone and moment material for the agent to draw on. Keep material consistent with the knowledge base to avoid price/policy conflicts.
维护可由 AI 主动引荐给客户的真人专属顾问名片库，录入引荐条件、审核与启停。仅在账号显式开启「辅助模式」时启用——这是全自治红线的受控例外。
A library of human advisor cards the AI can proactively introduce, with referral conditions, review and on/off. Active only when an account opts into "assist mode" — a controlled exception to the autonomy line.
管理后台总控 Agent、方法论生成 Agent 与跨模块 Prompt Pack。提示词分层（灵魂→系统契约→策略→业务上下文→操作指令）并版本化，运行日志记录所用版本。
Manage the management agent, the playbook-generation agent and the cross-module prompt pack. Prompts are layered (Soul → System Contract → Policy → Business Context → Operator) and versioned; run logs record which version was used.
管理 LLM 服务商：base_url / api_key / model / 协议格式，兼容主流 Chat Completions 与 Messages 协议。支持测试连通性与一键热切换激活配置，无需重启。
Manage LLM providers: base_url / api_key / model / protocol, compatible with mainstream Chat Completions and Messages APIs. Test connectivity and hot-swap the active config in one click — no restart.
实时监控自治回路：修订触发率、AI 暂缓三类细分（策略主动暂缓 / 安全门拦截 / 等待更多上下文）、未验证产品声明拦截、发送链路状态与最近修订记录。
Monitor the autonomy loop live: rewrite trigger rate, the three hold categories (policy hold / safety block / awaiting context), unverified-claim blocks, send-path status and recent revisions.
查看自演化器产出的 experiments、阈值与 Prompt 候选、Shadow 影子评测与显著性结论。影子重放只读历史决策、零副作用；显著性达标且通过安全回归门，管理员二次确认后才发布，可一键回滚。
Review the evolver's experiments, threshold and prompt candidates, shadow evaluations and significance verdicts. Shadow replay only reads past decisions — zero side effects. Release only on significance and a passed safety gate, after admin confirmation; rollback in one click.
追踪跟进任务、Agent 决策事件与系统执行结果。每一次决策、每一次工具调用、每一次发送都留痕，构成完整的执行审计链。
Track follow-up tasks, agent decision events and execution results. Every decision, tool call and send is recorded, forming a complete execution audit trail.
用户回复率、对话深度等长期指标，知识切片自动校验，方法论公式遵守度评测，产品声明兜底标记词管理。把「运营得好不好」量化成可追踪的数据。
Long-term metrics like reply rate and conversation depth, automatic slice validation, methodology-compliance scoring and fallback claim-marker management — quantifying "how well it operates."
查看 AI 主动发送的素材与专属顾问名片的使用次数、覆盖客户数、响应率与阶段推进率，衡量主动触达与引荐动作的真实效果。
See usage counts, customers covered, response and stage-advance rates for proactively sent material and advisor cards — measuring the real effect of outreach and referrals.
AI 永远不会替你做超出职权的决定。它把所有需要决策源拍板的事项，按 8 个清晰的待办流汇聚到同一收件箱，逐条处置后再由 AI 用自己的口吻执行。
The AI never makes calls beyond its mandate. It funnels everything that needs a decision-maker into one inbox across 8 clear streams; once handled, the AI executes in its own voice.
超出职权的事项向幕后决策源请示
Consult the principal on out-of-mandate matters
AI 起草的知识需人工核验才生效
AI-drafted knowledge needs human verification
AI 自创标签待审，未审不进决策
AI-proposed tags await review before use
关系类型识别结果待确认
Confirm detected relationship types
行业画像变更发布前二次确认
Confirm domain-profile changes before publish
自演化候选达标后人工发布 / 回滚
Release or roll back evolution candidates
沉淀的经验晋升为可复用知识候选
Promote lessons into reusable knowledge
对话暴露的知识盲区待补全
Fill knowledge gaps surfaced in conversation
18 个频道不是 18 个孤岛。它们共享同一个发送网关、同一套评审闸门、同一份可审计日志。
18 channels are not 18 silos. They share one send gateway, one set of review gates, and one auditable log.